AI output is already addressed in copyright law. It’s covered, in part, under derivative work and parody provisions, as well as other provisions. In other words, under existing copyright law, all AI output is, and can be, copyrighted by the one who generated the output. Furthermore, copyright is automatic in these cases, especially in cases when the method of generation is not specifically stated.
What I mean is that in the same way a physical media artist is not required to disclose what brand of brushes, paint, canvas, or reference material they used to create their art, a digital artist is not required to disclose the software applications, plugins, or generative tools they used to create their art. This is because the output is always considered either an original work or a derivative work, depending on whether there was reference material used, and the copyright belongs to the individual who generated the output. The tools themselves might be copyrighted, but the output from those tools is copyrighted by those who create the output.
For example, I use Photoshop and Illustrator and a number of plugins and tools, along with stock photos and images, and even AI to create my art. The output from these tools is copyrighted by me automatically under the law because I was the individual who used the tools that generated the output on my computer and into a file, just like a painter used the tools of paint, brushes, and physical motions to generate output upon a canvas. The creator of the tool has no rights to my work.
The only catch is that the final output must differ significantly from the original reference material. Roblox explains this quite well in their article on Intellectual Property. When you’re using known source material, it’s always a good idea to ask for written permission from the original artist if possible. If it’s not possible, make sure your art is significantly and sufficiently different. The great thing about AI generated art is that most output usually is.
The confusion and debate about copyright with AI stems from the material used to train the models. Under US copyright laws regarding compilations and collective works, the models compiled copyrighted works without the explicit permission of the original copyright holders. However, because of the nature of an AI model, the compilation structure (organization, file format, etc.) constitutes a derivative work, which is not illegal. The output using that model is also a derivative work as it is based on new information created by the model from information contained in the original material. That makes the AI model itself technically covered by copyright, but the creators of many of the models rarely seek copyright protection, if they ever do. This causes the AI models to be considered “fair use” or “public domain” under the law, as the model then meets the four required criteria to be considered fair use. It also puts the output from these models at least two levels of derivative work away from the original source material, further solidifying the output as a derivative work eligible for copyright by the individual who generated it.
To illustrate all this, consider this workflow:
flat PNG file ← layered Photoshop file with multiple edits ← Photoshop application ← AI output ← AI application ← AI Model ← AI model trainer ← original Source images
The original source images are copyrighted, but the trainer used them under fair use (AI research, non-profit) to create both a collective and derivative work that anyone can freely use. That model is then used freely by an AI application to generate a new, derivative output (JPG, GIF, or PNG) by another individual that then uses Photoshop to create another derivative work and saved in a new file. In this example, there are three levels of derivative work; the AI model itself, the AI output from the model, and the Photoshop output.
Now all that said, the US Copyright office said that AI generated art falls into the public domain. This means it’s not subject to copyright. The case that brought this decision is getting appealed, but it does solidify my main point, which is that AI art is a derivative work. Their decision also explicitly states that any significant editing done by image editing software such as Photoshop constitutes a new work which CAN be copyrighted by the individual that did it. It also solidifies that AI generated art does NOT violate the copyright of the original source images used to train the models.